CN106204615A - Salient target detection method based on central rectangular composition prior - Google Patents

Abstract

The invention provides a significant target detection method based on central rectangular composition prior. The central rectangle is a rectangle surrounded by four composition intersections of three composition lines. Supposing that the salient objects are arranged along the central rectangular composition line, performing relevance sequencing on the superpixels on four sides of the central rectangular to obtain a central rectangular composition line salient; supposing that the salient object is positioned on the central rectangular composition intersection point, removing composition intersection points which cannot become the salient object according to the central rectangular composition line salient map, then respectively taking the remaining composition intersection points as central nodes, calculating the spatial distances between all super-pixel nodes and the central nodes in the image to form corresponding salient maps, and finally adding and fusing the corresponding salient maps to form a central rectangular composition intersection point salient map; then, a compactness relation saliency map is obtained by utilizing the compactness relation; and finally, fusing the three to obtain a final saliency map. The method follows the law of photography and composition and conforms to the human eye visual attention mechanism.

Description

A kind of well-marked target detection method based on central rectangular composition priori

Technical field

The invention belongs to computer vision field, be specifically related to a kind of well-marked target detection method.

Background technology

Well-marked target in computer vision detection in recent years increasingly attracts the concern of people.Well-marked target detection is multiplex In the work such as image segmentation, target recognition, video tracking, image classification, compression of images, belong to the basis in computer vision Research work.Research worker it is also proposed the algorithm of a lot of relevant well-marked target detection.

Yang in 2013 et al. is at paper Saliency Detection via Graph-Based Manifold Ranking proposes MR method, image is carried out super-pixel segmentation, the super-pixel node at the four edges circle place of image is set to Background, finds target background scattergram according to feature correlation sequence, then sets out with the target searched out, feature correlation arrange The notable figure of sequence refinement.This is a kind of method forming notable figure from background angle.

Owing to image generally follows photography composition rule when being formed, as photographer is when shooting image, by the master of image Body target is placed on composition intersection point, or along the patterned lines multiple targets of arrangement.And, human eye, when watching image, also can be abided by Follow the photography composition rule of correspondence.Application number 201510402217.X mono-kind multiple dimensioned well-marked target detection side based on patterned lines Method, with patterned lines as target, remaining angle as background sets out, by repeatedly feature correlation sequence progressive updating target and the back of the body Scape, forms notable figure.But in patterned lines, the part of close image boundary is not likely to be target, if the target of being initialized with Noise can be produced.

Zhou in 2015 et al. is at paper Salient Region Detection via Integrating Diffusion-Based Compactness and Local Contrast proposes Compactness method, it is believed that notable Target has compact spatial distribution, and the color of background is broadly distributed in whole image, and thinks that well-marked target is many It is in the center of image, less with the space length of picture centre.But based on photography composition rule, well-marked target differs It is scheduled on the center of image, and is distributed across on composition intersection point, using the space length with picture centre as the foundation significantly calculated, Error can be produced.

Summary of the invention

The present invention is the patterned lines in the method overcoming and calculating image saliency value with patterned lines for target near image boundary It is initialized as noise that target brought and assumes that picture centre is error problem produced by well-marked target, it is provided that Yi Zhongji Well-marked target detection method in central rectangular composition priori.Described central rectangular refers to four composition intersection points of three points of patterned lines The rectangle surrounded.Assume that well-marked target arranges along central rectangular patterned lines, the super-pixel on central rectangular four edges is carried out spy Levy relevance ranking, obtain central rectangular patterned lines and significantly scheme；Then, it is assumed that well-marked target is positioned at central rectangular composition intersection point On, by calculating in image the space length between super-pixel node and Centroid, obtain central rectangular composition intersection point notable Figure, and significantly scheme according to compactedness Relation acquisition compactedness relation；Finally, three is merged and is obtained final notable figure.

The present invention solves technical problem and adopts the following technical scheme that

A kind of well-marked target detection method based on central rectangular composition priori, its step includes:

(1) divide the image into as super-pixel, with super-pixel as node, structure closed loop figure；

(2) assuming that target arranges along central rectangular patterned lines, the super-pixel node extracting central rectangular patterned lines place is made For query node, calculated the saliency value of each super-pixel node by relevance ranking, obtain central rectangular patterned lines notable Figure；

(3) assume that target is positioned on central rectangular composition intersection point, significantly scheme to remove not according to described central rectangular patterned lines It is likely to become the composition intersection point of well-marked target, the most respectively node centered by remaining composition intersection point, calculates in image all Super-pixel node and the space length of Centroid, form corresponding notable figure, be finally added fusion and form central rectangular Composition intersection point is significantly schemed；When eliminating all of composition intersection point according to the described notable figure of central rectangular patterned lines, with in image Node centered by the super-pixel node at heart place；

(4) utilize spaces compact sexual relationship to calculate the saliency value of image, obtain compactedness relation and significantly scheme；

(5) described central rectangular patterned lines significantly schemed, central rectangular composition intersection point is significantly schemed and compactedness relation is notable Figure three is merged, and obtains final central rectangular composition priori and significantly schemes.

Compared with the prior art, the present invention has the beneficial effect that:

1, a kind of well-marked target detection method based on central rectangular composition priori of the present invention, it is assumed that central rectangular patterned lines Form notable figure for well-marked target, it then follows photography composition rule, meet human eye vision attention mechanism；

2, a kind of well-marked target detection method based on central rectangular composition priori of the present invention, it is assumed that during well-marked target is positioned at On heart rectangle composition intersection point rather than image center location, it then follows photography composition rule, meet human eye vision attention mechanism；

3, a kind of well-marked target detection method based on central rectangular composition priori of the present invention, by image library test comparison Demonstrate its effectiveness and in effect obvious advantage.

Accompanying drawing explanation

Fig. 1 is a kind of well-marked target detection method flow chart based on central rectangular composition priori of the present invention.

Fig. 2 is super-pixel structure closed loop in a kind of well-marked target detection method based on central rectangular composition priori of the present invention The schematic diagram of figure.

Fig. 3 is a kind of well-marked target detection method based on central rectangular composition priori of the present invention with existing method in data Significance testing result PR curve comparison figure on collection CSSD.

Fig. 4 is a kind of well-marked target detection method based on central rectangular composition priori of the present invention with existing method in data Significance testing result PR curve comparison figure on collection ECSSD.

Fig. 5 is a kind of well-marked target detection method based on central rectangular composition priori of the present invention with existing method in data Significance testing result PR curve comparison figure on collection THUS-10000.

Fig. 6 is a kind of well-marked target detection method based on central rectangular composition priori of the present invention with existing method in data The significance histogrammic comparison diagram of testing result evaluation index on collection CSSD.

Fig. 7 is a kind of well-marked target detection method based on central rectangular composition priori of the present invention with existing method in data The significance histogrammic comparison diagram of testing result evaluation index on collection ECSSD.

Fig. 8 is a kind of well-marked target detection method based on central rectangular composition priori of the present invention with existing method in data The significance histogrammic comparison diagram of testing result evaluation index on collection THUS-10000.

Below by way of detailed description of the invention, and the present invention will be further described to combine accompanying drawing, but the embodiment party of the present invention Formula is not limited to this.

Detailed description of the invention

A kind of well-marked target detection method based on central rectangular composition priori of the present embodiment, as it is shown in figure 1, its step bag Include:

(1) utilize SLIC algorithm to divide the image into as super-pixel, with super-pixel as node, each node is set not only and week Enclose neighbor node to be connected, and be connected with the node on all borders altogether.Meanwhile, the node on central rectangular patterned lines four edges All regard as adjoining, the node on image four edges circle is also regarded as adjoining, constructs closed loop figure.As shown in Figure 2.

(2) assuming that target arranges along central rectangular patterned lines, the super-pixel node extracting central rectangular patterned lines place is made For query node, use manifold ranking algorithm to calculate the saliency value of each super-pixel node, obtain central rectangular patterned lines and show Write figure；

The four edges set of central rectangular is respectively left margin E_l, right margin E_r, coboundary E_tWith lower boundary E_d:

$E_l=(\[\frac{WIDTH}{3}\],\[\frac{HEIGHT}{3},...,\frac{2*HEIGHT}{3}\])---\left(1\right)$

$E_r=(\[\frac{2*WIDTH}{3}\],\[\frac{HEIGHT}{3},...,\frac{2*HEIGHT}{3}\])---\left(2\right)$

$E_t=(\[\frac{WIDTH}{3},...,\frac{2*WIDTH}{3}\],\[\frac{HEIGHT}{3}\])---\left(3\right)$

$E_d=(\[\frac{WIDTH}{3},...,\frac{2*WIDTH}{3}\],\[\frac{2*HEIGHT}{3}\])---\left(4\right)$

Wherein, HEIGHT and WIDTH represents the height and width of image respectively." [] " expression rounds.

Then, using manifold ranking algorithm to be ranked up super-pixel node, ranking functions is:

g^*=(D-α W)^-1y (5)

$w_{ij}=\left\{\begin{array}{cc}\exp (-\frac{l_{ij}}{{\mathrm{\σ}}^2})& ifj\∈Ni\\ 0& otherwise\end{array}\right.---\left(6\right)$

d_ii=∑_jw_ij (7)

Wherein W=[w_ij]_N×NFor incidence matrix, l_ijRepresent node v_iWith node v_jBetween Lab color space distance, N_i Represent node v_iThe set of neighbor node.D=diag{d₁₁,…,d_nnIt is degree matrix, α=0.99 is coefficient factor,For instruction vector, Unique (x) function representation takes unduplicated data in x.

By result g after sequence^*I () normalization, obtains central rectangular patterned lines and significantly schemes So.

$So={\stackrel{\‾}{g}}^{*}\left(i\right)---\left(8\right)$

Using the well-marked target that obtains as query node, reuse manifold ranking algorithm and carry out dependency with other node Sequence, can obtain more accurate notable figure.

(3) assume that target is positioned on central rectangular composition intersection point, significantly scheme to remove not according to described central rectangular patterned lines It is likely to become the composition intersection point of well-marked target, the most respectively node centered by remaining composition intersection point, calculates in image all Super-pixel node and the space length of Centroid, form corresponding notable figure, be finally added fusion and form central rectangular Composition intersection point is significantly schemed；When eliminating all of composition intersection point according to the described notable figure of central rectangular patterned lines, with in image Node centered by the super-pixel node at heart place.

First, the similarity between super-pixel is calculated:

$a_{ij}=\exp (-\frac{l_{ij}}{{\mathrm{\σ}}^2})---\left(9\right)$

Then, similarity is propagated:

F^*T=(D-α W)^-1A (10)

Wherein, matrix A=[a_ij]_N×N, similar matrix after diffusion

Afterwards, calculate central rectangular composition intersection point and significantly scheme Sd (i).

$Sd\left(i\right)=\left\{\begin{array}{cc}\underset{p_k\∈I}{\overset{I}{\mathrm{\Σ}}}\frac{{\mathrm{\Σ}}_{j=1}^R{f^{*}}_{ij}\·n_j\·\left|\right|b_j-p_k\left|\right|}{{\mathrm{\Σ}}_{j=1}^R{f^{*}}_{ij}\·n_j}& p_k\∈I\\ \frac{{\mathrm{\Σ}}_{j=1}^R{f^{*}}_{ij}\·n_j\·\left|\right|b_j-p_0\left|\right|}{{\mathrm{\Σ}}_{j=1}^R{f^{*}}_{ij}\·n_j}& p_k\∉I\end{array}\right.---\left(11\right)$

Wherein, p_k=[p_kx,p_ky], k=1,2,3,4 represent the space coordinates of four composition intersection points, p₀=[p_0x,p_0y] table The space coordinates of diagram inconocenter,Represent super-pixel v_jBarycenter, n_jRepresent super-pixel v_jThe number of middle pixel, R Representing super-pixel number in whole image-region, I represents the center on the well-marked target that central rectangular patterned lines significantly schemes So Rectangle composition intersection point set.

(4) utilize spaces compact sexual relationship to calculate the significance of image, obtain compactedness relation and significantly scheme.

$Sv\left(i\right)=\frac{{\mathrm{\Σ}}_{j=1}^R{f^{*}}_{ij}\·n_j\·\left|\right|b_j-{\mathrm{\μ}}_i\left|\right|}{{\mathrm{\Σ}}_{j=1}^R{f^{*}}_{ij}\·n_j}---\left(12\right)$

Wherein, spatial mean valueIt is defined as:

${\mathrm{\μ}}_i^x=\frac{{\mathrm{\Σ}}_{j=1}^R{f^{*}}_{ij}\·n_j\·b_j^x}{{\mathrm{\Σ}}_{j=1}^R{f^{*}}_{ij}\·n_j}---\left(13\right)$

${\mathrm{\μ}}_i^y=\frac{{\mathrm{\Σ}}_{j=1}^R{f^{*}}_{ij}\·n_j\·b_j^y}{{\mathrm{\Σ}}_{j=1}^R{f^{*}}_{ij}\·n_j}---\left(14\right)$

(5) described central rectangular patterned lines significantly schemed, central rectangular composition intersection point is significantly schemed and compactedness relation is notable Figure three is merged, and obtains final central rectangular composition priori and significantly schemes.

S=Norm (So*exp (1-Norm (Sv (i)+Sd (i)))) (15)

Wherein, x is normalized by Norm (x) function representation.

A kind of well-marked target detection method based on central rectangular composition priori of the present embodiment, with central rectangular composition priori Knowledge is set out, and in conjunction with spaces compact sexual relationship, improves the accuracy rate of well-marked target detection.By at data set CSSD, ECSSD, Carrying out significance detection on THUS-10000, testing result PR curve ratio is relatively such as Fig. 3, shown in Fig. 4, Fig. 5, evaluation index rectangular histogram Relatively such as Fig. 6, shown in Fig. 7, Fig. 8, obtain good Detection results, absolutely prove effectiveness and the universality of method.

Claims (1)

1. a well-marked target detection method based on central rectangular composition priori, its step includes:

(1) divide the image into as super-pixel, with super-pixel as node, structure closed loop figure；

(2) assume that target arranges along central rectangular patterned lines, extract the super-pixel node at central rectangular patterned lines place as looking into Ask node, calculated the saliency value of each super-pixel node by relevance ranking, obtain central rectangular patterned lines and significantly scheme；

(3) assuming that target is positioned on central rectangular composition intersection point, it is impossible significantly to scheme to remove according to described central rectangular patterned lines Become the composition intersection point of well-marked target, the most respectively node centered by remaining composition intersection point, calculate all super pictures in image Element node and the space length of Centroid, form corresponding notable figure, be finally added fusion and form central rectangular composition Intersection point is significantly schemed；When eliminating all of composition intersection point according to the described notable figure of central rectangular patterned lines, with picture centre institute Super-pixel node centered by node；

(4) utilize spaces compact sexual relationship to calculate the saliency value of image, obtain compactedness relation and significantly scheme；

(5) described central rectangular patterned lines significantly schemed, central rectangular composition intersection point is significantly schemed and the notable figure of compactedness relation three Person is merged, and obtains final central rectangular composition priori and significantly schemes.